Bypass Model Router - Use Direct Model

Document Version: 1.0
Last Updated: December 31, 2025
Status: Troubleshooting Guide

Problem

The Model Router may be selecting a model (like o4-mini) that causes errors or requires specific API versions. You want to bypass the Model Router and use a direct model deployment (e.g., gpt-5.1-chat or gpt-5-chat) that worked before.

Solution: Disable Model Router

The Model Router is enabled when AZURE_AI_MODEL_ROUTER is set. To bypass it and use a direct model:

Step 1: Clear Model Router Environment Variable

In Azure Container Apps:

  1. Go to Azure Portal > Container Apps > Your app
  2. Go to Configuration > Environment variables
  3. Find AZURE_AI_MODEL_ROUTER
  4. Either:
    • Delete the variable, OR
    • Set it to empty string ""

Step 2: Set Direct Model Deployment

Ensure AZURE_AI_DEPLOYMENT is set to your direct model:

AZURE_AI_DEPLOYMENT="gpt-5.1-chat"  # or "gpt-5-chat" if that's what worked

In Azure Container Apps:

  1. Go to Configuration > Environment variables
  2. Set AZURE_AI_DEPLOYMENT to your direct model name (e.g., gpt-5.1-chat)

Step 3: Verify Configuration

After updating, the backend will:

  • Use AZURE_AI_DEPLOYMENT directly (not Model Router)
  • Log: "Using direct deployment: {deployment}" instead of "Using Model Router"
  • Call the model directly without routing

Configuration Comparison

With Model Router (Current - May Cause Issues) 

AZURE_AI_ENDPOINT="https://zimax-gw.azure-api.net/zimax"
AZURE_AI_DEPLOYMENT="model-router"  # Not used when MODEL_ROUTER is set
AZURE_AI_MODEL_ROUTER="model-router"  # ← This enables Model Router
AZURE_AI_API_VERSION="2024-12-01-preview"

Result: Model Router selects model (may choose o4-mini or other models)

Without Model Router (Direct Model - Recommended) 

AZURE_AI_ENDPOINT="https://zimax-gw.azure-api.net/zimax"
AZURE_AI_DEPLOYMENT="gpt-5.1-chat"  # ← Direct model name
AZURE_AI_MODEL_ROUTER=""  # ← Empty or not set (disables Model Router)
AZURE_AI_API_VERSION="2024-10-01-preview"  # Can use older version

Result: Direct call to gpt-5.1-chat model

Code Behavior

When AZURE_AI_MODEL_ROUTER is not set or empty:

# backend/agents/base.py
deployment = self.settings.azure_ai_deployment  # Uses direct model
# Model Router check is skipped
# Calls: {endpoint}/openai/deployments/gpt-5.1-chat/chat/completions

When AZURE_AI_MODEL_ROUTER is set:

# backend/agents/base.py
if self.settings.azure_ai_model_router:
    deployment = self.settings.azure_ai_model_router  # Uses Model Router
    # Model Router selects which model to use
    # Calls: {endpoint}/openai/deployments/model-router/chat/completions

Quick Fix (Azure Portal)

  1. Azure Portal > Container Apps > staging-env-api
  2. Configuration > Environment variables
  3. Edit AZURE_AI_MODEL_ROUTER:
    • Option A: Delete the variable
    • Option B: Set value to empty string ""
  4. Save (triggers new revision)
  5. Wait for deployment (~14 minutes)

Verification

After deployment, check backend logs for:

With Model Router (should NOT appear):

INFO: Using Model Router via APIM Gateway: model-router

Without Model Router (should appear):

INFO: FoundryChatClient: Calling {endpoint}/openai/deployments/gpt-5.1-chat/chat/completions

Testing

After disabling Model Router, test chat:

python3 scripts/test-chat-quick.py --token <JWT_TOKEN> --message "hi"

Expected: Chat should work with direct gpt-5.1-chat model.

  • docs/sop/azure-foundry-chat-sop.md - Model Router configuration
  • backend/agents/base.py - Model selection logic
  • backend/core/config.py - Configuration settings

Notes

  • Model Router is useful for automatic model selection based on workload
  • Direct Model gives you control over which model is used
  • If Model Router was selecting o4-mini, disabling it will use your configured AZURE_AI_DEPLOYMENT instead
  • You can re-enable Model Router later by setting AZURE_AI_MODEL_ROUTER again